AI/ ai · llm-safety · interpretability · research

Hidden Reasoning in LLMs Is More Readable Than You Think

Researchers decoded what frontier models compute during silent filler tokens, recovering intermediate reasoning steps with 80-95% accuracy without any labels.

AI models can reason without showing their work — and a new paper suggests we can read that work anyway.

Researchers tested two open-weights frontier models, DeepSeek V3 and Kimi K2, on tasks where the models produce correct answers using content-free filler tokens like dots or counting sequences instead of a visible chain-of-thought. Using attention maps, logit-lens readouts, and a technique called KV-cache transplants, the team traced exactly how the models routed reasoning through those silent positions. They then built an unsupervised decoding pipeline that takes only hidden states as input — no ground-truth labels, no fine-tuning — and recovered intermediate values with 80-95% accuracy across four task types: fact retrieval, numeric composition, string manipulation, and in-context computation.

The finding matters because filler-token reasoning has been treated as a blind spot for AI oversight: if a model produces an answer with no visible reasoning steps, behavioral monitoring has nothing to examine. This paper pushes back on that framing, arguing that monitorability is a property of the full computational trace, not just the tokens a model prints. That's a meaningful shift in how safety researchers might think about what's actually auditable inside a model.

The caveat is that this was demonstrated on specific, bounded tasks with two models — not a general proof that all hidden computation is recoverable. Whether the technique scales to longer reasoning chains or less structured problems is the next question nobody has answered yet.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →